[For maintainers] Suggested jobs to run (before merge)

run-slow: afmoe, aimv2, albert, align, altclip, apertus, arcee, aria, audio_spectrogram_transformer, audioflamingo3, auto, autoformer, aya_vision, bamba, bark, bart

I'm not very knowledgeable about mypy, but this seems redundant to me: we're already defining in each module's __all__ attribute which objects we want to export. If we want to add a new object to export, we now need to add it to the module's __all__ as well as every downstream __init__.py file which would import it, instead of the current wildcard approach.

For models specifically this was always a feature more than a bug: any object in the model module's __all__ should be propagated up to the main transformers init.

Since my issue has been mentioned here: From my personal point of view, I believe that import from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLForConditionalGeneration is no worse than from transformers import Qwen3VLForConditionalGeneration. Modern IDEs/editors can autocomplete these entries, so there's actually no need to import everything in __init__.py. (Of course, this might be a bit radical and lead to a lack of backward compatibility.)

There is an even more extreme idea: Since the transformers package itself is already too complex, I think it's necessary to break it down into multiple packages instead of keeping it in a single package. This would allow users to use model A without considering anything related to model B (including downloading and analyzing code that wouldn't be used at all - that's a waste of time and computing resources). (Of course, I'm going a bit overboard with this. I believe there are reasons for the current design.)